Commit ff648221 authored by Patrick von Platen's avatar Patrick von Platen
Browse files

fix conflicts

parent c0d9dd3b
...@@ -81,6 +81,7 @@ class PretrainedConfig(object): ...@@ -81,6 +81,7 @@ class PretrainedConfig(object):
self.pad_token_id = kwargs.pop("pad_token_id", None) self.pad_token_id = kwargs.pop("pad_token_id", None)
self.eos_token_ids = kwargs.pop("eos_token_ids", None) self.eos_token_ids = kwargs.pop("eos_token_ids", None)
self.length_penalty = kwargs.pop("length_penalty", 1.0) self.length_penalty = kwargs.pop("length_penalty", 1.0)
self.no_repeat_ngram_size = kwargs.pop("no_repeat_ngram_size", 0)
self.num_return_sequences = kwargs.pop("num_return_sequences", 1) self.num_return_sequences = kwargs.pop("num_return_sequences", 1)
# Fine-tuning task arguments # Fine-tuning task arguments
......
...@@ -16,7 +16,6 @@ ...@@ -16,7 +16,6 @@
import logging import logging
import math import math
import random import random
import ipdb
from typing import Dict, List, Optional, Tuple from typing import Dict, List, Optional, Tuple
import torch import torch
......
...@@ -173,7 +173,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -173,7 +173,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
if getattr(output_embeddings, "bias", None) is not None: if getattr(output_embeddings, "bias", None) is not None:
output_embeddings.bias.data = torch.nn.functional.pad( output_embeddings.bias.data = torch.nn.functional.pad(
output_embeddings.bias.data, output_embeddings.bias.data,
(0, output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0]), (0, output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],),
"constant", "constant",
0, 0,
) )
...@@ -411,7 +411,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -411,7 +411,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
else: else:
raise EnvironmentError( raise EnvironmentError(
"Error no file named {} found in directory {} or `from_tf` set to False".format( "Error no file named {} found in directory {} or `from_tf` set to False".format(
[WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME + ".index"], pretrained_model_name_or_path [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME + ".index",],
pretrained_model_name_or_path,
) )
) )
elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path): elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
...@@ -425,7 +426,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -425,7 +426,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
archive_file = pretrained_model_name_or_path + ".index" archive_file = pretrained_model_name_or_path + ".index"
else: else:
archive_file = hf_bucket_url( archive_file = hf_bucket_url(
pretrained_model_name_or_path, postfix=(TF2_WEIGHTS_NAME if from_tf else WEIGHTS_NAME) pretrained_model_name_or_path, postfix=(TF2_WEIGHTS_NAME if from_tf else WEIGHTS_NAME),
) )
# redirect to the cache, if necessary # redirect to the cache, if necessary
...@@ -520,7 +521,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -520,7 +521,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
def load(module: nn.Module, prefix=""): def load(module: nn.Module, prefix=""):
local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {}) local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
module._load_from_state_dict( module._load_from_state_dict(
state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs,
) )
for name, child in module._modules.items(): for name, child in module._modules.items():
if child is not None: if child is not None:
...@@ -620,6 +621,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -620,6 +621,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
pad_token_id=None, pad_token_id=None,
eos_token_ids=None, eos_token_ids=None,
length_penalty=None, length_penalty=None,
no_repeat_ngram_size=None,
num_return_sequences=None, num_return_sequences=None,
): ):
r""" Generates sequences for models with a LM head. The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling r""" Generates sequences for models with a LM head. The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling
...@@ -725,6 +727,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -725,6 +727,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
eos_token_ids = eos_token_ids if eos_token_ids is not None else self.config.eos_token_ids eos_token_ids = eos_token_ids if eos_token_ids is not None else self.config.eos_token_ids
length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty
no_repeat_ngram_size = (
no_repeat_ngram_size if no_repeat_ngram_size is not None else self.config.no_repeat_ngram_size
)
num_return_sequences = ( num_return_sequences = (
num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
) )
...@@ -754,6 +759,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -754,6 +759,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
isinstance(eos_token_ids, (list, tuple)) and ((isinstance(e, int) and e >= 0) for e in eos_token_ids) isinstance(eos_token_ids, (list, tuple)) and ((isinstance(e, int) and e >= 0) for e in eos_token_ids)
), "`eos_token_ids` should be a positive integer or a list/tuple of positive integers." ), "`eos_token_ids` should be a positive integer or a list/tuple of positive integers."
assert length_penalty > 0, "`length_penalty` should be strictly positive." assert length_penalty > 0, "`length_penalty` should be strictly positive."
assert (
isinstance(no_repeat_ngram_size, int) and no_repeat_ngram_size >= 0
), "`no_repeat_ngram_size` should be a positive integer."
assert ( assert (
isinstance(num_return_sequences, int) and num_return_sequences > 0 isinstance(num_return_sequences, int) and num_return_sequences > 0
), "`num_return_sequences` should be a strictly positive integer." ), "`num_return_sequences` should be a strictly positive integer."
...@@ -764,7 +772,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -764,7 +772,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
"or a `bos_token_id` (integer >= 0) as a first token to start the generation." "or a `bos_token_id` (integer >= 0) as a first token to start the generation."
) )
input_ids = torch.full( input_ids = torch.full(
(batch_size, 1), bos_token_id, dtype=torch.long, device=next(self.parameters()).device (batch_size, 1), bos_token_id, dtype=torch.long, device=next(self.parameters()).device,
) )
else: else:
assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)." assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."
...@@ -811,23 +819,17 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -811,23 +819,17 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
# TODO (PVP): check eos_token_id # TODO (PVP): check eos_token_id
# TODO (PVP): probably not the best way to check whether model is encoder decoder # TODO (PVP): probably not the best way to check whether model is encoder decoder
is_encoder_decoder = ( is_encoder_decoder = (
hasattr(self, "model") hasattr(self, "model") and hasattr(self.model, "decoder") and hasattr(self.model, "encoder")
and hasattr(self.model, "decoder")
and hasattr(self.model, "encoder")
) )
if is_encoder_decoder: if is_encoder_decoder:
eos_token_id = eos_token_ids[0] eos_token_id = eos_token_ids[0]
assert ( assert bos_token_id is not None, "Encoder Decoder Models need to have a bos_token_id"
bos_token_id is not None assert eos_token_id is not None, "Encoder Decoder Models need to have a eos_token_id"
), "Encoder Decoder Models need to have a bos_token_id"
assert (
eos_token_id is not None
), "Encoder Decoder Models need to have a eos_token_id"
# encoder decoder need to start with empty input_ids and copy the input_ids to encoder_inputs # encoder decoder need to start with empty input_ids and copy the input_ids to encoder_inputs
encoder_inputs = input_ids encoder_inputs = input_ids
input_ids = torch.full( input_ids = torch.full(
(effective_batch_size * num_beams, 1), (effective_batch_size * num_beams, 1),
# eos_token_id, # Why eos_token_id here? bos_token_id makes more sense no? # eos_token_id, # Why eos_token_id here? bos_token_id makes more sense no?
bos_token_id, bos_token_id,
dtype=torch.long, dtype=torch.long,
device=next(self.parameters()).device, device=next(self.parameters()).device,
...@@ -849,6 +851,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -849,6 +851,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
top_k, top_k,
top_p, top_p,
repetition_penalty, repetition_penalty,
no_repeat_ngram_size,
pad_token_id, pad_token_id,
eos_token_ids, eos_token_ids,
effective_batch_size, effective_batch_size,
...@@ -869,6 +872,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -869,6 +872,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
top_k, top_k,
top_p, top_p,
repetition_penalty, repetition_penalty,
no_repeat_ngram_size,
pad_token_id, pad_token_id,
eos_token_ids, eos_token_ids,
effective_batch_size, effective_batch_size,
...@@ -888,6 +892,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -888,6 +892,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
top_k, top_k,
top_p, top_p,
repetition_penalty, repetition_penalty,
no_repeat_ngram_size,
pad_token_id, pad_token_id,
eos_token_ids, eos_token_ids,
batch_size, batch_size,
...@@ -902,9 +907,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -902,9 +907,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
past = None past = None
while cur_len < max_length: while cur_len < max_length:
model_inputs = self.prepare_inputs_for_generation( model_inputs = self.prepare_inputs_for_generation(input_ids, past=past, encoder_inputs=encoder_inputs)
input_ids, past=past, encoder_inputs=encoder_inputs
)
outputs = self(**model_inputs) outputs = self(**model_inputs)
next_token_logits = outputs[0][:, -1, :] next_token_logits = outputs[0][:, -1, :]
...@@ -917,9 +920,20 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -917,9 +920,20 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
if repetition_penalty != 1.0: if repetition_penalty != 1.0:
self.enforce_repetition_penalty_(next_token_logits, batch_size, 1, input_ids, repetition_penalty) self.enforce_repetition_penalty_(next_token_logits, batch_size, 1, input_ids, repetition_penalty)
# calculate a list of banned tokens to prevent repetitively generating the same ngrams
if no_repeat_ngram_size > 0:
# from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
banned_tokens = calc_banned_tokens(input_ids, batch_size, no_repeat_ngram_size, cur_len)
for batch_idx in range(batch_size):
next_token_logits[
batch_idx, banned_tokens[batch_idx]
] = -10000.0 # set eos token prob to 0 as is done for attention masks
if eos_token_ids is not None and cur_len < min_length: if eos_token_ids is not None and cur_len < min_length:
for eos_token_id in eos_token_ids: for eos_token_id in eos_token_ids:
next_token_logits[:, eos_token_id] = -10000.0 # set eos token prob to 0 as is done for attention masks next_token_logits[
:, eos_token_id
] = -10000.0 # set eos token prob to 0 as is done for attention masks
if do_sample: if do_sample:
# Temperature (higher temperature => more likely to sample low probability tokens) # Temperature (higher temperature => more likely to sample low probability tokens)
...@@ -981,6 +995,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -981,6 +995,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
top_k, top_k,
top_p, top_p,
repetition_penalty, repetition_penalty,
no_repeat_ngram_size,
pad_token_id, pad_token_id,
eos_token_ids, eos_token_ids,
batch_size, batch_size,
...@@ -993,9 +1008,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -993,9 +1008,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
""" Generate sequences for each example with beam search. """ Generate sequences for each example with beam search.
""" """
is_encoder_decoder = ( is_encoder_decoder = (
hasattr(self, "model") hasattr(self, "model") and hasattr(self.model, "decoder") and hasattr(self.model, "encoder")
and hasattr(self.model, "decoder")
and hasattr(self.model, "encoder")
) )
# generated hypotheses # generated hypotheses
...@@ -1017,9 +1030,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -1017,9 +1030,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
done = [False for _ in range(batch_size)] done = [False for _ in range(batch_size)]
while cur_len < max_length: while cur_len < max_length:
model_inputs = self.prepare_inputs_for_generation( model_inputs = self.prepare_inputs_for_generation(input_ids, past=past, encoder_inputs=encoder_inputs)
input_ids, past=past, encoder_inputs=encoder_inputs
)
outputs = self(**model_inputs) # (batch_size * num_beams, cur_len, vocab_size) outputs = self(**model_inputs) # (batch_size * num_beams, cur_len, vocab_size)
next_token_logits = outputs[0][:, -1, :] # (batch_size * num_beams, vocab_size) next_token_logits = outputs[0][:, -1, :] # (batch_size * num_beams, vocab_size)
...@@ -1030,12 +1041,23 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -1030,12 +1041,23 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
# repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858) # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
if repetition_penalty != 1.0: if repetition_penalty != 1.0:
self.enforce_repetition_penalty_( self.enforce_repetition_penalty_(
next_token_logits, batch_size, num_beams, input_ids, repetition_penalty next_token_logits, batch_size, num_beams, input_ids, repetition_penalty,
) )
# calculate a list of banned tokens to prevent repetitively generating the same ngrams
if no_repeat_ngram_size > 0:
# from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
banned_tokens = calc_banned_tokens(input_ids, batch_size, no_repeat_ngram_size, cur_len)
for batch_idx in range(batch_size):
next_token_logits[
batch_idx, banned_tokens[batch_idx]
] = -10000.0 # set eos token prob to 0 as is done for attention masks
if eos_token_ids is not None and cur_len < min_length: if eos_token_ids is not None and cur_len < min_length:
for eos_token_id in eos_token_ids: for eos_token_id in eos_token_ids:
next_token_logits[:, eos_token_id] = -10000.0 # set eos token prob to 0 as is done for attention masks next_token_logits[
:, eos_token_id
] = -10000.0 # set eos token prob to 0 as is done for attention masks
if do_sample: if do_sample:
# Temperature (higher temperature => more likely to sample low probability tokens) # Temperature (higher temperature => more likely to sample low probability tokens)
...@@ -1070,14 +1092,18 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -1070,14 +1092,18 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
# do greedy beam search # do greedy beam search
scores = F.log_softmax(next_token_logits, dim=-1) # (batch_size * num_beams, vocab_size) scores = F.log_softmax(next_token_logits, dim=-1) # (batch_size * num_beams, vocab_size)
if is_encoder_decoder: # TODO(PVP) to be refactored later if is_encoder_decoder: # TODO(PVP) to be refactored later - do we need this boolean flag here?
# scores[scores != scores] = -math.inf # block nans => seems very hacky here # scores[scores != scores] = -math.inf # block nans => seems very hacky here
# scores[:, pad_token_id] = -math.inf => seems very hacky here # scores[:, pad_token_id] = -math.inf => seems very hacky here
# TODO(SS): fairseq also takes out <unk> every step, and has unk at slot 3 # TODO(SS): fairseq also takes out <unk> every step, and has unk at slot 3
# if cur_len == 0: # Force BOS to be chosen => also very hacky ... seems also to work without this line # if cur_len == 0: # Force BOS to be chosen => also very hacky ... seems also to work without this line
# scores[:, self.config.bos_token_id + 1 :] = -math.inf # scores[:, self.config.bos_token_id + 1 :] = -math.inf
if cur_len == max_length - 1: # FORCE EOS to be chosen if cur_len == max_length - 1: # FORCE EOS to be chosen
all_but_eos_mask = torch.tensor([x for x in range(vocab_size) if x not in eos_token_ids], dtype=torch.long, device=next(self.parameters()).device) all_but_eos_mask = torch.tensor(
[x for x in range(vocab_size) if x not in eos_token_ids],
dtype=torch.long,
device=next(self.parameters()).device,
)
scores[:, all_but_eos_mask] = -10000.0 scores[:, all_but_eos_mask] = -10000.0
assert scores.size() == (batch_size * num_beams, vocab_size) assert scores.size() == (batch_size * num_beams, vocab_size)
...@@ -1175,7 +1201,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -1175,7 +1201,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
assert torch.all( assert torch.all(
next_scores[batch_idx, :num_beams] == beam_scores.view(batch_size, num_beams)[batch_idx] next_scores[batch_idx, :num_beams] == beam_scores.view(batch_size, num_beams)[batch_idx]
), "If batch_idx is not done, final next scores: {} have to equal to accumulated beam_scores: {}".format( ), "If batch_idx is not done, final next scores: {} have to equal to accumulated beam_scores: {}".format(
next_scores[:, :num_beams][batch_idx], beam_scores.view(batch_size, num_beams)[batch_idx] next_scores[:, :num_beams][batch_idx], beam_scores.view(batch_size, num_beams)[batch_idx],
) )
# need to add best num_beams hypotheses to generated hyps # need to add best num_beams hypotheses to generated hyps
...@@ -1218,7 +1244,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -1218,7 +1244,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
assert (len(hypo) == max_length for hypo in best) assert (len(hypo) == max_length for hypo in best)
decoded = torch.stack(best).type(torch.long).to(next(self.parameters()).device) decoded = torch.stack(best).type(torch.long).to(next(self.parameters()).device)
return decoded[:, 1:] if is_encoder_decoder:
# do not return first <BOS> token
return decoded[:, 1:]
return decoded
@staticmethod @staticmethod
def _reorder_cache(past, beam_idx): def _reorder_cache(past, beam_idx):
...@@ -1235,6 +1264,30 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin): ...@@ -1235,6 +1264,30 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
return past return past
def calc_banned_tokens(prev_input_ids, num_hypos, no_repeat_ngram_size, step):
# Copied from fairseq for no_repeat_ngram in beam_search"""
if step + 2 < no_repeat_ngram_size:
# return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
return [[] for _ in range(num_hypos)]
generated_ngrams = [{} for _ in range(num_hypos)]
for idx in range(num_hypos):
gen_tokens = prev_input_ids[idx].tolist()
generated_ngram = generated_ngrams[idx]
for ngram in zip(*[gen_tokens[i:] for i in range(no_repeat_ngram_size)]):
prev_ngram_tuple = tuple(ngram[:-1])
generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
def _get_generated_ngrams(hypo_idx):
# Before decoding the next token, prevent decoding of ngrams that have already appeared
start_idx = step + 2 - no_repeat_ngram_size
end_idx = step + 1
ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:end_idx].tolist())
return generated_ngrams[hypo_idx].get(ngram_idx, [])
banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
return banned_tokens
def top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1): def top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1):
""" Filter a distribution of logits using top-k and/or nucleus (top-p) filtering """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
Args: Args:
...@@ -1508,7 +1561,7 @@ class SQuADHead(nn.Module): ...@@ -1508,7 +1561,7 @@ class SQuADHead(nn.Module):
self.answer_class = PoolerAnswerClass(config) self.answer_class = PoolerAnswerClass(config)
def forward( def forward(
self, hidden_states, start_positions=None, end_positions=None, cls_index=None, is_impossible=None, p_mask=None self, hidden_states, start_positions=None, end_positions=None, cls_index=None, is_impossible=None, p_mask=None,
): ):
outputs = () outputs = ()
...@@ -1567,7 +1620,7 @@ class SQuADHead(nn.Module): ...@@ -1567,7 +1620,7 @@ class SQuADHead(nn.Module):
start_states = torch.einsum("blh,bl->bh", hidden_states, start_log_probs) start_states = torch.einsum("blh,bl->bh", hidden_states, start_log_probs)
cls_logits = self.answer_class(hidden_states, start_states=start_states, cls_index=cls_index) cls_logits = self.answer_class(hidden_states, start_states=start_states, cls_index=cls_index)
outputs = (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits) + outputs outputs = (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits,) + outputs
# return start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits # return start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits
# or (if labels are provided) (total_loss,) # or (if labels are provided) (total_loss,)
...@@ -1636,7 +1689,7 @@ class SequenceSummary(nn.Module): ...@@ -1636,7 +1689,7 @@ class SequenceSummary(nn.Module):
output = hidden_states.mean(dim=1) output = hidden_states.mean(dim=1)
elif self.summary_type == "cls_index": elif self.summary_type == "cls_index":
if cls_index is None: if cls_index is None:
cls_index = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2] - 1, dtype=torch.long) cls_index = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2] - 1, dtype=torch.long,)
else: else:
cls_index = cls_index.unsqueeze(-1).unsqueeze(-1) cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),)) cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
......
...@@ -16,7 +16,6 @@ ...@@ -16,7 +16,6 @@
import tempfile import tempfile
import unittest import unittest
import ipdb
from transformers import is_torch_available from transformers import is_torch_available
...@@ -87,7 +86,7 @@ class ModelTester: ...@@ -87,7 +86,7 @@ class ModelTester:
eos_token_ids=self.eos_token_id, eos_token_ids=self.eos_token_id,
bos_token_id=self.bos_token_id, bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id, pad_token_id=self.pad_token_id,
output_past=self.output_past output_past=self.output_past,
) )
inputs_dict = prepare_bart_inputs_dict(config, input_ids) inputs_dict = prepare_bart_inputs_dict(config, input_ids)
return config, inputs_dict return config, inputs_dict
...@@ -107,14 +106,8 @@ def prepare_bart_inputs_dict( ...@@ -107,14 +106,8 @@ def prepare_bart_inputs_dict(
@require_torch @require_torch
class BARTModelTest(ModelTesterMixin, unittest.TestCase): class BARTModelTest(ModelTesterMixin, unittest.TestCase):
<<<<<<< HEAD all_model_classes = (BartModel, BartForConditionalGeneration, BartForSequenceClassification) if is_torch_available() else ()
all_model_classes = ( all_generative_model_classes = (BartForConditionalGeneration,) if is_torch_available() else ()
(BartModel, BartForConditionalGeneration, BartForSequenceClassification) if is_torch_available() else ()
)
=======
all_model_classes = (BartModel, BartForMaskedLM, BartForSequenceClassification) if is_torch_available() else ()
all_generative_model_classes = (BartForMaskedLM,) if is_torch_available() else ()
>>>>>>> add to pass first tests
is_encoder_decoder = True is_encoder_decoder = True
# TODO(SS): fix the below in a separate PR # TODO(SS): fix the below in a separate PR
test_pruning = False test_pruning = False
...@@ -126,10 +119,10 @@ class BARTModelTest(ModelTesterMixin, unittest.TestCase): ...@@ -126,10 +119,10 @@ class BARTModelTest(ModelTesterMixin, unittest.TestCase):
self.model_tester = ModelTester(self) self.model_tester = ModelTester(self)
self.config_tester = ConfigTester(self, config_class=BartConfig) self.config_tester = ConfigTester(self, config_class=BartConfig)
def _A_test_config(self): def test_config(self):
self.config_tester.run_common_tests() self.config_tester.run_common_tests()
def _A_test_advanced_inputs(self): def test_advanced_inputs(self):
# (config, input_ids, token_type_ids, input_mask, *unused) = \ # (config, input_ids, token_type_ids, input_mask, *unused) = \
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
decoder_input_ids, decoder_attn_mask = _prepare_bart_decoder_inputs(config, inputs_dict["input_ids"]) decoder_input_ids, decoder_attn_mask = _prepare_bart_decoder_inputs(config, inputs_dict["input_ids"])
...@@ -168,7 +161,7 @@ class BARTModelTest(ModelTesterMixin, unittest.TestCase): ...@@ -168,7 +161,7 @@ class BARTModelTest(ModelTesterMixin, unittest.TestCase):
)[0] )[0]
_assert_tensors_equal(decoder_features_with_long_encoder_mask, decoder_features_with_created_mask) _assert_tensors_equal(decoder_features_with_long_encoder_mask, decoder_features_with_created_mask)
def _A_test_save_load_strict(self): def test_save_load_strict(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
model = model_class(config) model = model_class(config)
...@@ -224,7 +217,7 @@ class BartHeadTests(unittest.TestCase): ...@@ -224,7 +217,7 @@ class BartHeadTests(unittest.TestCase):
) )
return config, input_ids, batch_size return config, input_ids, batch_size
def _A_test_sequence_classification_forward(self): def test_sequence_classification_forward(self):
config, input_ids, batch_size = self._get_config_and_data() config, input_ids, batch_size = self._get_config_and_data()
labels = _long_tensor([2] * batch_size).to(torch_device) labels = _long_tensor([2] * batch_size).to(torch_device)
model = BartForSequenceClassification(config) model = BartForSequenceClassification(config)
...@@ -236,7 +229,7 @@ class BartHeadTests(unittest.TestCase): ...@@ -236,7 +229,7 @@ class BartHeadTests(unittest.TestCase):
loss = outputs[0] loss = outputs[0]
self.assertIsInstance(loss.item(), float) self.assertIsInstance(loss.item(), float)
def _A_test_lm_forward(self): def test_lm_forward(self):
config, input_ids, batch_size = self._get_config_and_data(output_past=False) config, input_ids, batch_size = self._get_config_and_data(output_past=False)
decoder_lm_labels = ids_tensor([batch_size, input_ids.shape[1]], self.vocab_size).to(torch_device) decoder_lm_labels = ids_tensor([batch_size, input_ids.shape[1]], self.vocab_size).to(torch_device)
lm_model = BartForConditionalGeneration(config) lm_model = BartForConditionalGeneration(config)
...@@ -248,7 +241,7 @@ class BartHeadTests(unittest.TestCase): ...@@ -248,7 +241,7 @@ class BartHeadTests(unittest.TestCase):
self.assertEqual(logits.shape, expected_shape) self.assertEqual(logits.shape, expected_shape)
self.assertIsInstance(loss.item(), float) self.assertIsInstance(loss.item(), float)
def _A_test_lm_uneven_forward(self): def test_lm_uneven_forward(self):
config = BartConfig( config = BartConfig(
vocab_size=self.vocab_size, vocab_size=self.vocab_size,
d_model=24, d_model=24,
...@@ -267,13 +260,8 @@ class BartHeadTests(unittest.TestCase): ...@@ -267,13 +260,8 @@ class BartHeadTests(unittest.TestCase):
expected_shape = (*summary.shape, config.vocab_size) expected_shape = (*summary.shape, config.vocab_size)
self.assertEqual(logits.shape, expected_shape) self.assertEqual(logits.shape, expected_shape)
<<<<<<< HEAD
def test_generate_beam_search(self): def test_generate_beam_search(self):
input_ids = torch.Tensor([[71, 82, 2], [68, 34, 2]]).long().to(torch_device) input_ids = torch.Tensor([[71, 82, 2], [68, 34, 2]]).long().to(torch_device)
=======
def _A_test_generate_beam_search(self):
input_ids = torch.Tensor([[71, 82, 2], [68, 34, 2]]).long()
>>>>>>> add to pass first tests
config = BartConfig( config = BartConfig(
vocab_size=self.vocab_size, vocab_size=self.vocab_size,
d_model=24, d_model=24,
...@@ -287,21 +275,19 @@ class BartHeadTests(unittest.TestCase): ...@@ -287,21 +275,19 @@ class BartHeadTests(unittest.TestCase):
output_past=True, output_past=True,
eos_token_ids=2, eos_token_ids=2,
pad_token_id=1, pad_token_id=1,
bos_token_id=0 bos_token_id=0,
) )
lm_model = BartForConditionalGeneration(config).to(torch_device) lm_model = BartForConditionalGeneration(config).to(torch_device)
lm_model.eval() lm_model.eval()
# new_input_ids = lm_model.generate( max_length = 5
# input_ids.clone(), num_return_sequences=1, num_beams=2, no_repeat_ngram_size=3, max_length=5
# )
new_input_ids = lm_model.generate( new_input_ids = lm_model.generate(
input_ids.clone(), num_return_sequences=1, num_beams=2, max_length=5 input_ids.clone(), num_return_sequences=1, num_beams=2, no_repeat_ngram_size=3, max_length=max_length
) )
self.assertEqual(new_input_ids.shape, (input_ids.shape[0], 5)) self.assertEqual(new_input_ids.shape, (input_ids.shape[0], max_length - 1))
# TODO(SS): uneven length batches, empty inputs # TODO(SS): uneven length batches, empty inputs
def _A_test_shift_tokens_right(self): def test_shift_tokens_right(self):
input_ids = torch.Tensor([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]]).long() input_ids = torch.Tensor([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]]).long()
shifted = shift_tokens_right(input_ids, 1) shifted = shift_tokens_right(input_ids, 1)
n_pad_before = input_ids.eq(1).float().sum() n_pad_before = input_ids.eq(1).float().sum()
...@@ -311,7 +297,7 @@ class BartHeadTests(unittest.TestCase): ...@@ -311,7 +297,7 @@ class BartHeadTests(unittest.TestCase):
self.assertTrue(torch.eq(shifted[:, 0], 2).all()) self.assertTrue(torch.eq(shifted[:, 0], 2).all())
@slow @slow
def _A_test_tokenization(self): def test_tokenization(self):
tokenizer = BartTokenizer.from_pretrained("bart-large") tokenizer = BartTokenizer.from_pretrained("bart-large")
examples = [" Hello world", " DomDramg"] # need leading spaces for equality examples = [" Hello world", " DomDramg"] # need leading spaces for equality
fairseq_results = [ fairseq_results = [
...@@ -387,7 +373,7 @@ TOLERANCE = 1e-4 ...@@ -387,7 +373,7 @@ TOLERANCE = 1e-4
@require_torch @require_torch
class BartModelIntegrationTest(unittest.TestCase): class BartModelIntegrationTest(unittest.TestCase):
@slow @slow
def _A_test_inference_no_head(self): def test_inference_no_head(self):
model = BartModel.from_pretrained("bart-large").to(torch_device) model = BartModel.from_pretrained("bart-large").to(torch_device)
input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]]) input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
inputs_dict = prepare_bart_inputs_dict(model.config, input_ids) inputs_dict = prepare_bart_inputs_dict(model.config, input_ids)
...@@ -401,7 +387,7 @@ class BartModelIntegrationTest(unittest.TestCase): ...@@ -401,7 +387,7 @@ class BartModelIntegrationTest(unittest.TestCase):
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=TOLERANCE)) self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=TOLERANCE))
@slow @slow
def _A_test_mnli_inference(self): def test_mnli_inference(self):
example_b = [0, 31414, 232, 328, 740, 1140, 69, 46078, 1588, 2, 1] example_b = [0, 31414, 232, 328, 740, 1140, 69, 46078, 1588, 2, 1]
input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2], example_b]) input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2], example_b])
...@@ -428,30 +414,34 @@ class BartModelIntegrationTest(unittest.TestCase): ...@@ -428,30 +414,34 @@ class BartModelIntegrationTest(unittest.TestCase):
_assert_tensors_equal(expected_slice, logits_arr, atol=TOLERANCE) _assert_tensors_equal(expected_slice, logits_arr, atol=TOLERANCE)
@unittest.skip("This is just too slow") @unittest.skip("This is just too slow")
def _A_test_model_from_pretrained(self): def test_model_from_pretrained(self):
# Forces 1.6GB download from S3 for each model # Forces 1.6GB download from S3 for each model
for model_name in list(BART_PRETRAINED_MODEL_ARCHIVE_MAP.keys()): for model_name in list(BART_PRETRAINED_MODEL_ARCHIVE_MAP.keys()):
model = BartModel.from_pretrained(model_name, cache_dir=CACHE_DIR) model = BartModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model) self.assertIsNotNone(model)
@slow @slow
def test_cnn_summarization_same_as_fairseq(self): def test_cnn_summarization_same_as_fairseq_easy(self):
hf = BartForConditionalGeneration.from_pretrained("bart-large-cnn", output_past=True,).to(torch_device) hf = BartForConditionalGeneration.from_pretrained("bart-large-cnn", output_past=True,).to(torch_device)
tok = BartTokenizer.from_pretrained("bart-large") tok = BartTokenizer.from_pretrained("bart-large")
text = " (CNN)The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday, a step that gives the court jurisdiction over alleged crimes in Palestinian" text = " (CNN)The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday, a step that gives the court jurisdiction over alleged crimes in Palestinian"
tokens = tok.encode(text, return_tensors="pt").to(torch_device) tokens = tok.encode(text, return_tensors="pt").to(torch_device)
extra_len = 20 extra_len = 20
gen_tokens_1 = hf.generate_1(tokens, num_beams=4, max_length=extra_len,) # repetition_penalty=10., gen_tokens_bart = hf.generate_1(tokens, num_beams=4, max_length=extra_len,) # repetition_penalty=10.,
gen_tokens = hf.generate(tokens, num_beams=4, max_length=extra_len + 2, do_sample=False) # repetition_penalty=10., gen_tokens = hf.generate(
print("1: {}".format(gen_tokens_1)) tokens, num_beams=4, max_length=extra_len + 2, do_sample=False
print("2: {}".format(gen_tokens)) ) # repetition_penalty=10.,
ipdb.set_trace()
expected_result = "<s>The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday." expected_result = "<s>The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday."
generated_1 = [tok.decode(g,) for g in gen_tokens_1] generated_bart = [tok.decode(g,) for g in gen_tokens_bart]
generated = [tok.decode(g,) for g in gen_tokens] generated = [tok.decode(g,) for g in gen_tokens]
self.assertEqual(expected_result, generated_bart[0])
self.assertEqual(expected_result, generated[0]) self.assertEqual(expected_result, generated[0])
# Harder cases with batching @slow
def test_cnn_summarization_same_as_fairseq_hard(self):
hf = BartForConditionalGeneration.from_pretrained("bart-large-cnn", output_past=True,).to(torch_device)
tok = BartTokenizer.from_pretrained("bart-large")
FRANCE_ARTICLE = ' Marseille, France (CNN)The French prosecutor leading an investigation into the crash of Germanwings Flight 9525 insisted Wednesday that he was not aware of any video footage from on board the plane. Marseille prosecutor Brice Robin told CNN that "so far no videos were used in the crash investigation." He added, "A person who has such a video needs to immediately give it to the investigators." Robin\'s comments follow claims by two magazines, German daily Bild and French Paris Match, of a cell phone video showing the harrowing final seconds from on board Germanwings Flight 9525 as it crashed into the French Alps. All 150 on board were killed. Paris Match and Bild reported that the video was recovered from a phone at the wreckage site. The two publications described the supposed video, but did not post it on their websites. The publications said that they watched the video, which was found by a source close to the investigation. "One can hear cries of \'My God\' in several languages," Paris Match reported. "Metallic banging can also be heard more than three times, perhaps of the pilot trying to open the cockpit door with a heavy object. Towards the end, after a heavy shake, stronger than the others, the screaming intensifies. Then nothing." "It is a very disturbing scene," said Julian Reichelt, editor-in-chief of Bild online. An official with France\'s accident investigation agency, the BEA, said the agency is not aware of any such video. Lt. Col. Jean-Marc Menichini, a French Gendarmerie spokesman in charge of communications on rescue efforts around the Germanwings crash site, told CNN that the reports were "completely wrong" and "unwarranted." Cell phones have been collected at the site, he said, but that they "hadn\'t been exploited yet." Menichini said he believed the cell phones would need to be sent to the Criminal Research Institute in Rosny sous-Bois, near Paris, in order to be analyzed by specialized technicians working hand-in-hand with investigators. But none of the cell phones found so far have been sent to the institute, Menichini said. Asked whether staff involved in the search could have leaked a memory card to the media, Menichini answered with a categorical "no." Reichelt told "Erin Burnett: Outfront" that he had watched the video and stood by the report, saying Bild and Paris Match are "very confident" that the clip is real. He noted that investigators only revealed they\'d recovered cell phones from the crash site after Bild and Paris Match published their reports. "That is something we did not know before. ... Overall we can say many things of the investigation weren\'t revealed by the investigation at the beginning," he said. What was mental state of Germanwings co-pilot? German airline Lufthansa confirmed Tuesday that co-pilot Andreas Lubitz had battled depression years before he took the controls of Germanwings Flight 9525, which he\'s accused of deliberately crashing last week in the French Alps. Lubitz told his Lufthansa flight training school in 2009 that he had a "previous episode of severe depression," the airline said Tuesday. Email correspondence between Lubitz and the school discovered in an internal investigation, Lufthansa said, included medical documents he submitted in connection with resuming his flight training. The announcement indicates that Lufthansa, the parent company of Germanwings, knew of Lubitz\'s battle with depression, allowed him to continue training and ultimately put him in the cockpit. Lufthansa, whose CEO Carsten Spohr previously said Lubitz was 100% fit to fly, described its statement Tuesday as a "swift and seamless clarification" and said it was sharing the information and documents -- including training and medical records -- with public prosecutors. Spohr traveled to the crash site Wednesday, where recovery teams have been working for the past week to recover human remains and plane debris scattered across a steep mountainside. He saw the crisis center set up in Seyne-les-Alpes, laid a wreath in the village of Le Vernet, closer to the crash site, where grieving families have left flowers at a simple stone memorial. Menichini told CNN late Tuesday that no visible human remains were left at the site but recovery teams would keep searching. French President Francois Hollande, speaking Tuesday, said that it should be possible to identify all the victims using DNA analysis by the end of the week, sooner than authorities had previously suggested. In the meantime, the recovery of the victims\' personal belongings will start Wednesday, Menichini said. Among those personal belongings could be more cell phones belonging to the 144 passengers and six crew on board. Check out the latest from our correspondents . The details about Lubitz\'s correspondence with the flight school during his training were among several developments as investigators continued to delve into what caused the crash and Lubitz\'s possible motive for downing the jet. A Lufthansa spokesperson told CNN on Tuesday that Lubitz had a valid medical certificate, had passed all his examinations and "held all the licenses required." Earlier, a spokesman for the prosecutor\'s office in Dusseldorf, Christoph Kumpa, said medical records reveal Lubitz suffered from suicidal tendencies at some point before his aviation career and underwent psychotherapy before he got his pilot\'s license. Kumpa emphasized there\'s no evidence suggesting Lubitz was suicidal or acting aggressively before the crash. Investigators are looking into whether Lubitz feared his medical condition would cause him to lose his pilot\'s license, a European government official briefed on the investigation told CNN on Tuesday. While flying was "a big part of his life," the source said, it\'s only one theory being considered. Another source, a law enforcement official briefed on the investigation, also told CNN that authorities believe the primary motive for Lubitz to bring down the plane was that he feared he would not be allowed to fly because of his medical problems. Lubitz\'s girlfriend told investigators he had seen an eye doctor and a neuropsychologist, both of whom deemed him unfit to work recently and concluded he had psychological issues, the European government official said. But no matter what details emerge about his previous mental health struggles, there\'s more to the story, said Brian Russell, a forensic psychologist. "Psychology can explain why somebody would turn rage inward on themselves about the fact that maybe they weren\'t going to keep doing their job and they\'re upset about that and so they\'re suicidal," he said. "But there is no mental illness that explains why somebody then feels entitled to also take that rage and turn it outward on 149 other people who had nothing to do with the person\'s problems." Germanwings crash compensation: What we know . Who was the captain of Germanwings Flight 9525? CNN\'s Margot Haddad reported from Marseille and Pamela Brown from Dusseldorf, while Laura Smith-Spark wrote from London. CNN\'s Frederik Pleitgen, Pamela Boykoff, Antonia Mortensen, Sandrine Amiel and Anna-Maja Rappard contributed to this report.' # @noqa FRANCE_ARTICLE = ' Marseille, France (CNN)The French prosecutor leading an investigation into the crash of Germanwings Flight 9525 insisted Wednesday that he was not aware of any video footage from on board the plane. Marseille prosecutor Brice Robin told CNN that "so far no videos were used in the crash investigation." He added, "A person who has such a video needs to immediately give it to the investigators." Robin\'s comments follow claims by two magazines, German daily Bild and French Paris Match, of a cell phone video showing the harrowing final seconds from on board Germanwings Flight 9525 as it crashed into the French Alps. All 150 on board were killed. Paris Match and Bild reported that the video was recovered from a phone at the wreckage site. The two publications described the supposed video, but did not post it on their websites. The publications said that they watched the video, which was found by a source close to the investigation. "One can hear cries of \'My God\' in several languages," Paris Match reported. "Metallic banging can also be heard more than three times, perhaps of the pilot trying to open the cockpit door with a heavy object. Towards the end, after a heavy shake, stronger than the others, the screaming intensifies. Then nothing." "It is a very disturbing scene," said Julian Reichelt, editor-in-chief of Bild online. An official with France\'s accident investigation agency, the BEA, said the agency is not aware of any such video. Lt. Col. Jean-Marc Menichini, a French Gendarmerie spokesman in charge of communications on rescue efforts around the Germanwings crash site, told CNN that the reports were "completely wrong" and "unwarranted." Cell phones have been collected at the site, he said, but that they "hadn\'t been exploited yet." Menichini said he believed the cell phones would need to be sent to the Criminal Research Institute in Rosny sous-Bois, near Paris, in order to be analyzed by specialized technicians working hand-in-hand with investigators. But none of the cell phones found so far have been sent to the institute, Menichini said. Asked whether staff involved in the search could have leaked a memory card to the media, Menichini answered with a categorical "no." Reichelt told "Erin Burnett: Outfront" that he had watched the video and stood by the report, saying Bild and Paris Match are "very confident" that the clip is real. He noted that investigators only revealed they\'d recovered cell phones from the crash site after Bild and Paris Match published their reports. "That is something we did not know before. ... Overall we can say many things of the investigation weren\'t revealed by the investigation at the beginning," he said. What was mental state of Germanwings co-pilot? German airline Lufthansa confirmed Tuesday that co-pilot Andreas Lubitz had battled depression years before he took the controls of Germanwings Flight 9525, which he\'s accused of deliberately crashing last week in the French Alps. Lubitz told his Lufthansa flight training school in 2009 that he had a "previous episode of severe depression," the airline said Tuesday. Email correspondence between Lubitz and the school discovered in an internal investigation, Lufthansa said, included medical documents he submitted in connection with resuming his flight training. The announcement indicates that Lufthansa, the parent company of Germanwings, knew of Lubitz\'s battle with depression, allowed him to continue training and ultimately put him in the cockpit. Lufthansa, whose CEO Carsten Spohr previously said Lubitz was 100% fit to fly, described its statement Tuesday as a "swift and seamless clarification" and said it was sharing the information and documents -- including training and medical records -- with public prosecutors. Spohr traveled to the crash site Wednesday, where recovery teams have been working for the past week to recover human remains and plane debris scattered across a steep mountainside. He saw the crisis center set up in Seyne-les-Alpes, laid a wreath in the village of Le Vernet, closer to the crash site, where grieving families have left flowers at a simple stone memorial. Menichini told CNN late Tuesday that no visible human remains were left at the site but recovery teams would keep searching. French President Francois Hollande, speaking Tuesday, said that it should be possible to identify all the victims using DNA analysis by the end of the week, sooner than authorities had previously suggested. In the meantime, the recovery of the victims\' personal belongings will start Wednesday, Menichini said. Among those personal belongings could be more cell phones belonging to the 144 passengers and six crew on board. Check out the latest from our correspondents . The details about Lubitz\'s correspondence with the flight school during his training were among several developments as investigators continued to delve into what caused the crash and Lubitz\'s possible motive for downing the jet. A Lufthansa spokesperson told CNN on Tuesday that Lubitz had a valid medical certificate, had passed all his examinations and "held all the licenses required." Earlier, a spokesman for the prosecutor\'s office in Dusseldorf, Christoph Kumpa, said medical records reveal Lubitz suffered from suicidal tendencies at some point before his aviation career and underwent psychotherapy before he got his pilot\'s license. Kumpa emphasized there\'s no evidence suggesting Lubitz was suicidal or acting aggressively before the crash. Investigators are looking into whether Lubitz feared his medical condition would cause him to lose his pilot\'s license, a European government official briefed on the investigation told CNN on Tuesday. While flying was "a big part of his life," the source said, it\'s only one theory being considered. Another source, a law enforcement official briefed on the investigation, also told CNN that authorities believe the primary motive for Lubitz to bring down the plane was that he feared he would not be allowed to fly because of his medical problems. Lubitz\'s girlfriend told investigators he had seen an eye doctor and a neuropsychologist, both of whom deemed him unfit to work recently and concluded he had psychological issues, the European government official said. But no matter what details emerge about his previous mental health struggles, there\'s more to the story, said Brian Russell, a forensic psychologist. "Psychology can explain why somebody would turn rage inward on themselves about the fact that maybe they weren\'t going to keep doing their job and they\'re upset about that and so they\'re suicidal," he said. "But there is no mental illness that explains why somebody then feels entitled to also take that rage and turn it outward on 149 other people who had nothing to do with the person\'s problems." Germanwings crash compensation: What we know . Who was the captain of Germanwings Flight 9525? CNN\'s Margot Haddad reported from Marseille and Pamela Brown from Dusseldorf, while Laura Smith-Spark wrote from London. CNN\'s Frederik Pleitgen, Pamela Boykoff, Antonia Mortensen, Sandrine Amiel and Anna-Maja Rappard contributed to this report.' # @noqa
EXPECTED_SUMMARY_FRANCE = 'French prosecutor says he\'s not aware of any video footage from on board the plane. German daily Bild and French Paris Match claim to have found a cell phone video of the crash. A French Gendarmerie spokesman calls the reports "completely wrong" and "unwarranted" German airline Lufthansa confirms co-pilot Andreas Lubitz had battled depression.' EXPECTED_SUMMARY_FRANCE = 'French prosecutor says he\'s not aware of any video footage from on board the plane. German daily Bild and French Paris Match claim to have found a cell phone video of the crash. A French Gendarmerie spokesman calls the reports "completely wrong" and "unwarranted" German airline Lufthansa confirms co-pilot Andreas Lubitz had battled depression.'
...@@ -465,28 +455,28 @@ class BartModelIntegrationTest(unittest.TestCase): ...@@ -465,28 +455,28 @@ class BartModelIntegrationTest(unittest.TestCase):
ARTICLE_SUBWAY = ' New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County, New York. A year later, she got married again in Westchester County, but to a different man and without divorcing her first husband. Only 18 days after that marriage, she got hitched yet again. Then, Barrientos declared "I do" five more times, sometimes only within two weeks of each other. In 2010, she married once more, this time in the Bronx. In an application for a marriage license, she stated it was her "first and only" marriage. Barrientos, now 39, is facing two criminal counts of "offering a false instrument for filing in the first degree," referring to her false statements on the 2010 marriage license application, according to court documents. Prosecutors said the marriages were part of an immigration scam. On Friday, she pleaded not guilty at State Supreme Court in the Bronx, according to her attorney, Christopher Wright, who declined to comment further. After leaving court, Barrientos was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the New York subway through an emergency exit, said Detective Annette Markowski, a police spokeswoman. In total, Barrientos has been married 10 times, with nine of her marriages occurring between 1999 and 2002. All occurred either in Westchester County, Long Island, New Jersey or the Bronx. She is believed to still be married to four men, and at one time, she was married to eight men at once, prosecutors say. Prosecutors said the immigration scam involved some of her husbands, who filed for permanent residence status shortly after the marriages. Any divorces happened only after such filings were approved. It was unclear whether any of the men will be prosecuted. The case was referred to the Bronx District Attorney\'s Office by Immigration and Customs Enforcement and the Department of Homeland Security\'s Investigation Division. Seven of the men are from so-called "red-flagged" countries, including Egypt, Turkey, Georgia, Pakistan and Mali. Her eighth husband, Rashid Rajput, was deported in 2006 to his native Pakistan after an investigation by the Joint Terrorism Task Force. If convicted, Barrientos faces up to four years in prison. Her next court appearance is scheduled for May 18.' ARTICLE_SUBWAY = ' New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County, New York. A year later, she got married again in Westchester County, but to a different man and without divorcing her first husband. Only 18 days after that marriage, she got hitched yet again. Then, Barrientos declared "I do" five more times, sometimes only within two weeks of each other. In 2010, she married once more, this time in the Bronx. In an application for a marriage license, she stated it was her "first and only" marriage. Barrientos, now 39, is facing two criminal counts of "offering a false instrument for filing in the first degree," referring to her false statements on the 2010 marriage license application, according to court documents. Prosecutors said the marriages were part of an immigration scam. On Friday, she pleaded not guilty at State Supreme Court in the Bronx, according to her attorney, Christopher Wright, who declined to comment further. After leaving court, Barrientos was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the New York subway through an emergency exit, said Detective Annette Markowski, a police spokeswoman. In total, Barrientos has been married 10 times, with nine of her marriages occurring between 1999 and 2002. All occurred either in Westchester County, Long Island, New Jersey or the Bronx. She is believed to still be married to four men, and at one time, she was married to eight men at once, prosecutors say. Prosecutors said the immigration scam involved some of her husbands, who filed for permanent residence status shortly after the marriages. Any divorces happened only after such filings were approved. It was unclear whether any of the men will be prosecuted. The case was referred to the Bronx District Attorney\'s Office by Immigration and Customs Enforcement and the Department of Homeland Security\'s Investigation Division. Seven of the men are from so-called "red-flagged" countries, including Egypt, Turkey, Georgia, Pakistan and Mali. Her eighth husband, Rashid Rajput, was deported in 2006 to his native Pakistan after an investigation by the Joint Terrorism Task Force. If convicted, Barrientos faces up to four years in prison. Her next court appearance is scheduled for May 18.'
EXPECTED_SUMMARY_SUBWAY = "Liana Barrientos has been married 10 times, sometimes within two weeks of each other. Prosecutors say the marriages were part of an immigration scam. On Friday, she pleaded not guilty at State Supreme Court in the Bronx. She was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the subway." EXPECTED_SUMMARY_SUBWAY = "Liana Barrientos has been married 10 times, sometimes within two weeks of each other. Prosecutors say the marriages were part of an immigration scam. On Friday, she pleaded not guilty at State Supreme Court in the Bronx. She was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the subway."
# dct = tok.batch_encode_plus( dct = tok.batch_encode_plus(
# [FRANCE_ARTICLE, SHORTER_ARTICLE, IRAN_ARTICLE, ARTICLE_SUBWAY], [FRANCE_ARTICLE, SHORTER_ARTICLE, IRAN_ARTICLE, ARTICLE_SUBWAY],
# max_length=1024, max_length=1024,
# pad_to_max_length=True, pad_to_max_length=True,
# return_tensors="pt", return_tensors="pt",
# ) )
# self.assertEqual(1024, dct["input_ids"].shape[1]) self.assertEqual(1024, dct["input_ids"].shape[1])
# hypotheses_batch = hf.generate( hypotheses_batch = hf.generate(
# input_ids=dct["input_ids"].to(torch_device), input_ids=dct["input_ids"].to(torch_device),
# attention_mask=dct["attention_mask"].to(torch_device), attention_mask=dct["attention_mask"].to(torch_device),
# num_beams=4, num_beams=4,
# length_penalty=2.0, length_penalty=2.0,
# max_length=140, max_length=140,
# min_len=55, min_len=55,
# no_repeat_ngram_size=3, no_repeat_ngram_size=3,
# ) )
# decoded = [ decoded = [
# tok.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in hypotheses_batch tok.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in hypotheses_batch
# ] ]
# self.assertListEqual( self.assertListEqual(
# [EXPECTED_SUMMARY_FRANCE, EXPECTED_SUMMARY_SHORTER, EXPECTED_SUMMARY_IRAN, EXPECTED_SUMMARY_SUBWAY], [EXPECTED_SUMMARY_FRANCE, EXPECTED_SUMMARY_SHORTER, EXPECTED_SUMMARY_IRAN, EXPECTED_SUMMARY_SUBWAY],
# decoded, decoded,
# ) )
# TODO(SS): run fairseq again with num_beams=2, min_len=20. # TODO(SS): run fairseq again with num_beams=2, min_len=20.
# TODO(SS): add test case that hits max_length # TODO(SS): add test case that hits max_length
...@@ -54,13 +54,13 @@ class ModelTesterMixin: ...@@ -54,13 +54,13 @@ class ModelTesterMixin:
model_tester = None model_tester = None
all_model_classes = () all_model_classes = ()
all_generative_model_classes = () all_generative_model_classes = ()
_A_test_torchscript = True test_torchscript = True
_A_test_pruning = True test_pruning = True
_A_test_resize_embeddings = True test_resize_embeddings = True
_A_test_head_masking = True test_head_masking = True
is_encoder_decoder = False is_encoder_decoder = False
def _A_test_save_load(self): def test_save_load(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -85,7 +85,7 @@ class ModelTesterMixin: ...@@ -85,7 +85,7 @@ class ModelTesterMixin:
max_diff = np.amax(np.abs(out_1 - out_2)) max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5) self.assertLessEqual(max_diff, 1e-5)
def _A_test_initialization(self): def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config) configs_no_init = _config_zero_init(config)
...@@ -99,7 +99,7 @@ class ModelTesterMixin: ...@@ -99,7 +99,7 @@ class ModelTesterMixin:
msg="Parameter {} of model {} seems not properly initialized".format(name, model_class), msg="Parameter {} of model {} seems not properly initialized".format(name, model_class),
) )
def _A_test_determinism(self): def test_determinism(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -116,7 +116,7 @@ class ModelTesterMixin: ...@@ -116,7 +116,7 @@ class ModelTesterMixin:
max_diff = np.amax(np.abs(out_1 - out_2)) max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5) self.assertLessEqual(max_diff, 1e-5)
def _A_test_attention_outputs(self): def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
seq_len = getattr(self.model_tester, "seq_length", None) seq_len = getattr(self.model_tester, "seq_length", None)
decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len) decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
...@@ -179,25 +179,25 @@ class ModelTesterMixin: ...@@ -179,25 +179,25 @@ class ModelTesterMixin:
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length], [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
) )
def _A_test_torchscript(self): def test_torchscript(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
self._create_and_check_torchscript(config, inputs_dict) self._create_and_check_torchscript(config, inputs_dict)
def _A_test_torchscript_output_attentions(self): def test_torchscript_output_attentions(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_attentions = True config.output_attentions = True
self._create_and_check_torchscript(config, inputs_dict) self._create_and_check_torchscript(config, inputs_dict)
def _A_test_torchscript_output_hidden_state(self): def test_torchscript_output_hidden_state(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_hidden_states = True config.output_hidden_states = True
self._create_and_check_torchscript(config, inputs_dict) self._create_and_check_torchscript(config, inputs_dict)
def _create_and_check_torchscript(self, config, inputs_dict): def _create_and_check_torchscript(self, config, inputs_dict):
if not self._A_test_torchscript: if not self.test_torchscript:
return return
configs_no_init = _config_zero_init(config) # To be sure we have no Nan configs_no_init = _config_zero_init(config) # To be sure we have no Nan
...@@ -245,8 +245,8 @@ class ModelTesterMixin: ...@@ -245,8 +245,8 @@ class ModelTesterMixin:
self.assertTrue(models_equal) self.assertTrue(models_equal)
def _A_test_headmasking(self): def test_headmasking(self):
if not self._A_test_head_masking: if not self.test_head_masking:
return return
global_rng.seed(42) global_rng.seed(42)
...@@ -299,8 +299,8 @@ class ModelTesterMixin: ...@@ -299,8 +299,8 @@ class ModelTesterMixin:
self.assertAlmostEqual(attentions[-1][..., -2, :, :].flatten().sum().item(), 0.0) self.assertAlmostEqual(attentions[-1][..., -2, :, :].flatten().sum().item(), 0.0)
self.assertNotEqual(attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0) self.assertNotEqual(attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)
def _A_test_head_pruning(self): def test_head_pruning(self):
if not self._A_test_pruning: if not self.test_pruning:
return return
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -328,8 +328,8 @@ class ModelTesterMixin: ...@@ -328,8 +328,8 @@ class ModelTesterMixin:
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads) self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1) self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
def _A_test_head_pruning_save_load_from_pretrained(self): def test_head_pruning_save_load_from_pretrained(self):
if not self._A_test_pruning: if not self.test_pruning:
return return
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -361,8 +361,8 @@ class ModelTesterMixin: ...@@ -361,8 +361,8 @@ class ModelTesterMixin:
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads) self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1) self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
def _A_test_head_pruning_save_load_from_config_init(self): def test_head_pruning_save_load_from_config_init(self):
if not self._A_test_pruning: if not self.test_pruning:
return return
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -392,8 +392,8 @@ class ModelTesterMixin: ...@@ -392,8 +392,8 @@ class ModelTesterMixin:
self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads) self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1) self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
def _A_test_head_pruning_integration(self): def test_head_pruning_integration(self):
if not self._A_test_pruning: if not self.test_pruning:
return return
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -449,7 +449,7 @@ class ModelTesterMixin: ...@@ -449,7 +449,7 @@ class ModelTesterMixin:
self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]}) self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})
def _A_test_hidden_states_output(self): def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -474,9 +474,9 @@ class ModelTesterMixin: ...@@ -474,9 +474,9 @@ class ModelTesterMixin:
], ],
) )
def _A_test_resize_tokens_embeddings(self): def test_resize_tokens_embeddings(self):
(original_config, inputs_dict,) = self.model_tester.prepare_config_and_inputs_for_common() (original_config, inputs_dict,) = self.model_tester.prepare_config_and_inputs_for_common()
if not self._A_test_resize_embeddings: if not self.test_resize_embeddings:
return return
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -516,7 +516,7 @@ class ModelTesterMixin: ...@@ -516,7 +516,7 @@ class ModelTesterMixin:
self.assertTrue(models_equal) self.assertTrue(models_equal)
def _A_test_model_common_attributes(self): def test_model_common_attributes(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes: for model_class in self.all_model_classes:
...@@ -594,7 +594,7 @@ class ModelTesterMixin: ...@@ -594,7 +594,7 @@ class ModelTesterMixin:
# self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape) # self.assertTrue(model.transformer.wte.weight.shape, model.lm_head.weight.shape)
# self.assertTrue(check_same_values(model.transformer.wte, model.lm_head)) # self.assertTrue(check_same_values(model.transformer.wte, model.lm_head))
def _A_test_inputs_embeds(self): def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
if not self.is_encoder_decoder: if not self.is_encoder_decoder:
...@@ -711,7 +711,7 @@ def floats_tensor(shape, scale=1.0, rng=None, name=None): ...@@ -711,7 +711,7 @@ def floats_tensor(shape, scale=1.0, rng=None, name=None):
@require_torch @require_torch
class ModelUtilsTest(unittest.TestCase): class ModelUtilsTest(unittest.TestCase):
@slow @slow
def _A_test_model_from_pretrained(self): def test_model_from_pretrained(self):
logging.basicConfig(level=logging.INFO) logging.basicConfig(level=logging.INFO)
for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]: for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
config = BertConfig.from_pretrained(model_name) config = BertConfig.from_pretrained(model_name)
...@@ -736,7 +736,7 @@ class ModelUtilsTest(unittest.TestCase): ...@@ -736,7 +736,7 @@ class ModelUtilsTest(unittest.TestCase):
class UtilsFunctionsTest(unittest.TestCase): class UtilsFunctionsTest(unittest.TestCase):
# tests whether the top_k_top_p function behaves as expected # tests whether the top_k_top_p function behaves as expected
def _A_test_top_k_top_p_filtering(self): def test_top_k_top_p_filtering(self):
logits = torch.tensor( logits = torch.tensor(
[ [
[ [
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment